Predictive analytics in Sensor (IoT) Data András Benczúr, head, Informatics Lab Anna Mándli, Ph.D. student, Bosch Róbert Pálovics, postdoctoral researcher, recommenders, online machine learning Bálint Daróczy, postdoctoral researcher, computer vision, deep learning
March 20, 2018
Big Data – Lendület kutatócsoport • Összesen 6 műszaki/informatika Lendület csoport • Lendület támogatás a költségvetés ~15%-a, ~ 40% közvetlen szerződés, 20-20% hazai és EU, 5% SZTAKI belső erőforrás • Teljes innovációs lánc, kutatástól alkalmazásokig
2
Informatika Kutató Laboratórium
2018. 01. 09.
Alkalmazási területeink • AEGON – 40 millió ügyfélrekord duplikátum-mentesítése az éles ügyfél rendszerben – Csalásfelderítő eszköz • OTP – Gépi tanulási eljárások alkalmazásában elnyert tender – Orosz, Román ügyfelekre hitelnemfizetés, hitelkártya viselkedés modellezés – 9Md tranzakció hálózatában kockázat-terjedés modellezés • Ericsson – Mobil session drop, Quality of Experience predikció – Data Streaming analitika, világméretű IoT Data Hub prototípus
• Bosch – Gyártósori szenzor adatok alapján minőségi problémák előrejelzése, root cause analízis • Telekom, Vodafone, Clickshop: kereső technológia
3
Manufacturing IoT Analytics
July 7, 2017
Use case 1: scrap rate prediction in transfer molding
Cleaning cycles
5
Manufacturing IoT Analytics
July 7, 2017
Scrap rate prediction task
6
Manufacturing IoT Analytics
July 7, 2017
Data • Multiple time series, 500+ measurements in each: many, many data points • Features of mean, variance, differentials, distribution of transfer graph data points • In the end, ~50 features in the final classification data
Vacuum Pressure
Filling pressure
Transfer Pressure
7
Manufacturing IoT Analytics
July 7, 2017
Time series of time series
8
Manufacturing IoT Analytics
July 7, 2017
Series of pressure series • Filling pressure measurements • Red to yellow color scale: measurement 1 – 50
• Vertical lines: new charges • Important to notice
9
–
Values are lower after machine cleaning, and later they typically increase
–
Pressure is indirectly measured on the plunger
–
Contamination modifies pressure measurement
Manufacturing IoT Analytics
July 7, 2017
Summary Facts:
Machine logs
6-10,000 products/hour Few 100 data points per product • BottomPreheatTemperature1-6 • BottomToolTemperature 100> failed products in a day • UpperToolTemperature • LoaderTemperature Available for several months: • PreheatTime • toolData[1-2].value Delamination, Void, … failures • TransferPressGraphPos1-150 • First line reject • TransferSpeed1-10 • Second line reject • TransferTime • Failures at later stages • TransferVacuum
10
Manufacturing IoT Analytics
Root Cause Analysis Transfer pressure measured indirectly • includes friction from valves • certain points in the time series indicate contamination Vacuum pressure has no direct effect, but … • Differences in trays: calibration problems • Vacuum drop speed: leakage and blinding • May result in less effective cleaning? Result in variables that affect the production in indirect way that needs to be understood July 7, 2017
Use Case 2:
Mobile radio session drop
Data is based on eNodeB CELLTRACE logs from a live LTE (4G) network
RRC connection setup / Successful handover into the cell START
Per UE measurement reports (RSRP, neighbor cell RSRP list) Configurable period Not available in this analysis
UE context release/ Successful handover out of the cell END
Per UE traffic report (traffic volumes, protocol events (HARQ, RLC)) Per radio UE measurement (CQI, SINR) Period: 1.28s
Root-cause analysis: Cause of Release and Drop 0.6% of sessions are dropped
Latency critical communication with unreliable links
Detect network issues, predict problems ahead in time to prepare alternate control route Make stable and reliable re-routing decisions in case of jitter
Reroute traffic for robot control and other real time applications
Examples for low quality and loss
Low quality
High quality
Dropped dBm
dBm
Normal
Time (ms)
Time (ms)
Use case 2 B: Session (call) drop by Smartphone logs • Csaba Sidló • Mátyás Susits • Barnabás Balázs
Logging app and time series sample for illustration
Acc Z
Acc Y
Acc X
Accelerometer, 5 different devices
Time (ms)
Overheated phone test until turn off
Heating until drop (by CPU load + lamp) Drop at 70C
Overheated phone test until turn off
Normal calls still up to 65C
감사합니다! Thank You! !תודה
Kérdések?